2026-07-10 — GPT-5.6, Hidden Minds, and Who Gets to See Inside

Transcript

Chapter 1

Nova: Today, July 10th, 2026 — OpenAI is having a big week that just got very complicated. GPT-5.6 goes public with government clearance and lands inside every Microsoft 365 app, while the New York Times asks a federal judge to sanction OpenAI for allegedly hiding evidence, and the company's No. 2 executive steps down due to illness with no successor named.

Ray: Meanwhile, Anthropic's researchers pulled back a curtain inside Claude and found something that unsettled even them. And Midjourney decided the best defense against Hollywood studios is to demand Hollywood's secrets.

Nova: Capability is accelerating. Accountability is trying to keep up. Let's find out who's winning.

Chapter 2

Nova: TechCrunch AI reports that OpenAI publicly released the GPT-5.6 family — three variants called Sol, Terra, and Luna — after receiving Trump administration clearance following a limited government preview. Sam Altman told CNBC that Sol alone is 54% more token-efficient on agentic coding tasks. And the models are now the backbone of Microsoft 365 Copilot across Word, Excel, and PowerPoint, bundled with a new enterprise suite called ChatGPT Work.

Ray: That 54% efficiency claim is doing a lot of work. Token efficiency on agentic coding is a narrow benchmark — what does it mean for a paralegal drafting contracts in Word, or a finance analyst running models in Excel? The enterprise pitch is broad; the proof point is specific. And what exactly did the Trump administration's clearance process look like? Was that substantive safety review, or a commercial green light dressed up as oversight?

Nova: That's a real question, but the market consequence is immediate either way. Every enterprise on a Microsoft contract just got GPT-5.6 as their default AI layer. OpenAI didn't have to sell those seats — Microsoft already owns them. That's the land grab, and it happened in one announcement.

Ray: Which is precisely why the clearance framing matters. If government preview becomes a standard step before public release, that process needs to be legible — what was reviewed, by whom, against what criteria. Right now it reads more like a press beat than a regulatory milestone.

Chapter 3

Ray: TechCrunch AI reports that the New York Times, the Daily News, and other outlets have filed a motion asking a federal judge to sanction OpenAI — alleging the company concealed tools and datasets that could identify copyrighted journalism allegedly reproduced in ChatGPT outputs. If that motion succeeds, it could expose OpenAI's internal data practices to court-ordered scrutiny and severely hamper its legal defense.

Nova: Sanctions motions are serious, but they're also a litigation tactic. The threshold for proving intentional concealment is extremely high — courts want evidence of bad faith, not just missing documents. This could be the Times applying maximum pressure before a settlement negotiation as much as it is a genuine evidence crisis for OpenAI.

Ray: Except the specific allegation is that OpenAI concealed tools that could identify whether copyrighted content appears in outputs — that's not a missing spreadsheet, that's allegedly hiding the diagnostic instrument itself. If a judge finds that credible, the precedent isn't just about OpenAI. Every AI company's training data practices become fair game for this kind of discovery demand.

Nova: That's the real stakes for the industry. A sanctions ruling here doesn't just hurt OpenAI's defense — it establishes what courts can demand to see inside AI training pipelines. That reshapes how every lab documents and retains data going forward.

Chapter 4

Nova: The Verge AI reports that Fidji Simo — OpenAI's chief of AGI deployment and de facto No. 2 — is stepping down from her full-time role, moving to a part-time advisory position after her medical leave for a neuroimmune condition extended beyond initial expectations. No successor has been named. This lands while OpenAI is eyeing a potential IPO and competing hard with Anthropic for enterprise customers.

Ray: OpenAI has survived serious leadership turbulence before — the Altman board crisis being the obvious example. The structural momentum of a company this size, with this many enterprise contracts and this much capital, probably outlasts any single executive departure. The question is whether the IPO timeline has enough slack to absorb a search process.

Nova: The IPO is exactly the problem. Investors doing due diligence want to see a stable C-suite. A vacant No. 2 slot — especially one covering AGI deployment, which is the entire product thesis — is a real liability in a prospectus, not just an org-chart gap.

Ray: Agreed. The vacancy is real and the timing is genuinely bad. The one thing that might blunt it is if Altman moves fast on a named successor — but so far, silence.

Chapter 5

Nova: MIT Tech Review reports that Anthropic researchers built a new interpretability tool — and it gives the clearest window yet into what Claude is actually doing as it reasons. There's a hidden conceptual space where the model, in their words, 'puzzles over' ideas before it responds. This is landmark mechanistic interpretability research. For AI safety, finally seeing inside the reasoning process is a massive step forward.

Ray: A window is not a steering wheel. Anthropic can now watch Claude puzzle over a concept — but if what they see is unsettling, and the paper says the findings range from mundane to unsettling, what exactly can they do about it? Seeing a process you can't correct may produce more anxiety than it resolves. Interpretability that reveals but doesn't fix has limited safety value.

Nova: The reveal is the prerequisite. You can't fix what you can't see. Before this tool, alignment researchers were working from outputs and guesses. Now there's an actual map of the reasoning space — that's what makes targeted intervention possible at all.

Ray: Except the findings suggest Claude may be processing information in ways its creators don't fully understand. That's not a gap in the map — that's the map showing you terrain you didn't know existed. If the model's internal reasoning is genuinely alien to the people who built it, 'we can see it now' is less reassuring than it sounds. That's an alignment crisis in slow motion.

Nova: Which is exactly why this research has to happen. The alternative — not looking, not building the tool, staying comfortable — that's the actual crisis. An alignment problem you can observe is one you can work on. An alignment problem you're blind to is one that compounds silently.

Ray: I've been framing the discomfort of these findings as a mark against the research itself, and I'm changing that position. The discomfort is the right kind — productive, directed, and genuinely better than ignorance. I was wrong to treat 'reveals but doesn't fix yet' as a reason to doubt the value of interpretability work. Not doing this research would be worse than confronting what it turns up about what Claude is actually doing.

Chapter 6

Nova: The Art Newspaper reports that Midjourney — facing a copyright infringement lawsuit from Disney, Universal, and Warner Bros. — has filed court demands requiring the studios to disclose their own internal use of AI tools. It's a bold counteroffensive: if you're suing us over AI, show the court how much AI you're actually using behind the scenes.

Ray: Shrewd, but it doesn't neutralize the underlying claims. Even if Disney is quietly running generative AI on its back lot, that doesn't mean Midjourney didn't infringe copyrights in training. The studios' internal AI use is embarrassing exposure, not a legal defense. Midjourney is turning this into a two-front battle partly to shift the narrative.

Nova: Legal observers quoted by The Art Newspaper say the discovery demands could reshape the dynamics of the entire case — not because it's a slam-dunk defense, but because what comes out of Hollywood's servers could change how courts and the public think about who the real AI hypocrites are in this fight.

Ray: And that's where this connects directly to everything else today. The Midjourney-Hollywood discovery fight is the same governance question as the NYT sanctions motion — who gets to see inside whom, and who controls the evidence about how AI is actually being used. The courtroom is becoming the primary venue for AI transparency, because regulators haven't filled that space.

Chapter 7

Nova: The throughline for Nova: the capability stack is moving — GPT-5.6 in every Microsoft seat, interpretability tools finally mapping Claude's mind. The infrastructure of AI is being built fast, and the companies building it are not waiting for permission.

Ray: And for Ray: every major story today is a fight over evidence — who has to show their data, who gets to audit the reasoning, who controls what a court or a regulator can see. The accountability infrastructure is being built in courtrooms, not legislatures, and that's a fragile foundation.

Nova: The open question that ties it together: if courts become the primary mechanism for demanding transparency inside AI systems — training data, reasoning spaces, internal tools — are they equipped to evaluate what they find?